ideas for a postmodern harmonic analysis · numerical harmonic analysis group ideas for a...

Historical Viewpoint A Change of Paradigms! Time-Frequency Analysis and Modern Applications Good(?) reasons for measure theory Standard Spaces Generalized Stochastic Processes

Numerical Harmonic Analysis Group

Ideas for a Postmodern Harmonic Analysis

Hans G. Feichtingerwww.nuhag.eu

Muenich, September 2013Rupert Lasser Farewell Workshop

Mathematical Analysis and Applications

Hans G. Feichtinger www.nuhag.eu Ideas for a Postmodern Harmonic Analysis


Private Vorbemerkungen

What was the situation in Harmonic Analysis that we found in our“early days”:During the events in the 70-th (winterschools, conferences inOberwolfach etc.) we had

Elmar Thoma and Horst Leptin; later Kaniuth;

Hans Reiter

Andrej Hulanicki

Allessandro Figa-Talamanca;

Pierre Eymard

Antoine Derighetti

Hans G. Feichtinger Ideas for a Postmodern Harmonic Analysis



Topics of the early conferences:

amenability

Fourier Algebra A(G )

nilpotent groups, special groups

multiplier questions

ideals

BUT I do not remember any discussion about the “goals ofharmonic analysis”, it was more like a collection of deep andinteresting mathematical problems.As a visitor giving a survey talk on “analysis in general” JeanDieudonne declared abstract harmonic analysis “off-stream”.




Some comments made by Hans Reiter gave guidance about theunderstanding of the “essential problems” ([citations]):

We are studying convolution, hence the Banach algebra(L1(G ), ‖ · ‖1

)in order to understand its ideal structure

(because we are interested in spectral analysis);

We are studying Segal algebras because they are similar

The natural setting for Fourier Analysis are LCA groups!(A.Weil);

the existence of null-sets makes existence of the devil evident(hence use approximation arguments);




Presentation of ideas, or formulation of theories can be providedalso in quite different formats (think of courses in mathematics,physics or engineering):

1 heuristic;

2 strictly formal;

3 by examples;

4 conceptual; (e.g. integral)

5 computational recipes (integration methods);

6 by demonstrating applications;

7 standard research versus change of paradigms(Thomas Kuhn!)



Typology of Mathematicians

Research and publication of results is a human activity, hencethere are different typologies:

the BIRDS and the FROGS (Notices);

the theory builders;

the problem solvers;

the bridge builders;

the educators;

the communicators;

the perfectionists;

the systematic types;

the pioneers;

the masters of ceremony (closing up);

the output oriented ones;



Typology of Mathematicians

Research and publication of results is a human activity, hencethere are also different orientations:

applied mathematical work;

application driven research;

application oriented research

application motivated research;

pure/explorative research;

Of course the tools used and the level of precision reachedwill/may depend on the orientation of the researcher resp.the tradition in the community.



Historical Notes

Clearly engineers and applied people are more interested in“working code” and concrete algorithms with some justificationwhile people at the theoretical end may be happy with an existenceand uniqueness result.Given the ability to do extensive computations within short time,based on an elaborate theory and using efficient code I want toproclaim Postmodern Analysis as the (hopefully/inevitablyupcoming) period where all the knowledge is coming together andprovides constructive and realizable methods.



Some Heuristic Claims in Fourier Analysis

From the engineering point of view once can describe a TILS(translation invariant linear systems resp. operators) are justconvolution operators. This comes from the following heuristicconsideration, using Tx f (y) = f (y − x), noting thatTx (δ0) = deltax :

f =

∫G

f (x) δx dx =

∫G

Tx (δ0)f (x)dx ;

hence with µ := T (δ0), the impulse response we have:

Tf =

∫G

T (Tx (δ0))f (x)dx =

∫G

Tx (µ)f (x)dx = f ∗ µ,

so at the Fourier transform side we get the transfer function

Tf = µ · f .



The Scandal

The above reasoning is clearly too vague to be appreciated bymathematicians (and can be only “saved” using distributiontheory), and indeed allows for what I. Sandberg the scandal insystem theory: which consists in based on his (for abstractharmonic analyst not surprising) observation:There are non-trivial linear operators from Cb(Rd ) into itself,which commute with translations, but are not convolutionoperators! In fact, one can use on of those translation invariantmeans σ on Cb(Rd ) (since Rd is amenable) and map f to theconstant times σ(f ), but clearly it is zero on C0(Rd ) (they havelimit zero at infinity).Of course this is partially due to the nonseparability of Cb(Rd )and one simply has to take C0(Rd ) instead to save thesituation (see my course notes on the subject).



Some Important Tools in Fourier Analysis

This seems to be like a strong warning to work on save ground,and make sure that all the things that we do “exist in a propersense” and to avoid problems. It is like saying: you should notmake use of an electric device if it is not TUV approved.So what are the most important objects and tools that we need inFourier Analysis resp. in (abstract or applied) Harmonic Analysis:It is the Fourier transform, convolution, Poisson’s formula etc. ...Just recall that we need to take integrals

f (ω) =

∫Rd

f (t) · e−2πiω·t dt (1)

f ∗ g(x) =

∫Rd

g(x − y)f (x) dx . (2)




Therefore it appears completely natural (nowadays at least) to usethe Lebesgue integral as the perfect integral, resistent to limits(unlike the poor Riemann integral), and providing us with theBanach space L1(G ) resp. even the Hilbert space L2(G ), with itsscalar product and the possibility of proofing Plancherel’s theorem,from the engineering point of view the all important “energyconservation law” for the Fourier transform.So we became used to the Fourier transform as a an INTEGRALtransform (which is unitary, properly extended), turningconvolution into pointwise multiplication (and vice versa).But does that help when I discuss (e.g.) chirp signals of the formh(x) = e ix2

(linear chirps). They satisfy Fh = h and thusobviously they are harmless (bounded!) Fourier multipliers.But does h ∗ f (x) exists almost everywhere for f ∈ L2(R)?And also the inverse FT is not really an integral transform!?




Even if there is no other immediate use for Lp-space for pdifferentfrom p = 1, 2,∞ one can argue, that Lp-spaces are very naturalscale of spaces, nicely closed under complex interpolation(according to Riesz-Thorin).A sceptical person may ask, what it help a user, to know that agiven function is in L3.4, but let us avoid this discussion here.But what are the multipliers between Lp-spaces, or from Lp to Lq

(for general p, q ∈ [1,∞])?It turns out they are convolution operators with pseudo-measuresresp. quasi-measures (objects which are locally likepseudo-measures, the functionals on the Fourier algebra A(G ). Sothese limitations just shift the problem areas to other regions.There is also a (tricky) theory of transformable (possiblyunbounded) measures by Gil de Lamadrid.




Furthermore, let us not forget that we are going to do FOURIERANALYSIS and FOURIER SYNTHESIS originally, i.e. we wantedto filter out the frequencies inside a given signal using Fourieranalysis, and we would like to resynthesize the given signal fromthose frequencies which are found inside the signal, or it leastinside a set of frequencies where we can be sure (for whateverreasons) that it contains all the frequencies in the signal. In fact,this brings us again back to the original meaning of the concept ofsets of spectral synthesis!An it is not so difficult when we do everything just over finitegroup, or more down to earth, if we work with finite signals, viewedas functions on ZN or ZM × ZN in the case of images.



Abstract versus Conceptual Harmonic Analysis

Branches of Harmonic Analysis

AHA = Abstract Harmonic Analysis;

NHA = Numerical (or computational) Harmonic Analysis;

Applied Harmonic Analysis;

CONCEPTUAL HARMONIC ANALYSIS(integration of all this!)

I am going to give some verbal description of my view on AbstractHarmonic Analyis versus a term that I try to propagate more andmore, namely “Conceptual Harmonic Analysis”



Books on the market: Harmonic Analysis

Juergen Jost: Postmodern Analysis . 2nd ed. [19]Over the years the concept of Fourier Analysis and HarmonicAnalysis have changed very much.The “old testament” by Antoni Zygmund: Trigonometric series.2nd ed. Vols. I, II Cambridge University Press, (1959, [27])Edwin Hewitt and Kenneth A. Ross: Abstract HarmonicAnalysis. Vol. II: Structure and Analysis for Compact Groups.Analysis on Locally Compact Abelian Groups [18]Grafakos, Loukas: Classical and Modern Fourier Analysis [11]has later been split into two piecesHans Reiter [and Jan Stegeman]: “Classical HarmonicAnalysis and Locally Compact Groups” (2nd ed., 2000, [23]).



Books on the market

More recent books are:

H. J. Weaver: Applications of Discrete and ContinuousFourier Analysis (1983, [25]).

R. Shakarchi and Elias M. Stein: Fourier Analysis: AnIntroduction Princeton University Press, Princeton Lectures inAnalysis, (2003, [24]).

Robert Marks: Handbook of Fourier Analysis and itsApplications Oxford University Press, (2009 , [22]).

M.W. Wong: Discrete Fourier analysis. Pseudo-DifferentialOperators. Theory and Applications 5. Basel: Birkhauser(2011, [26]).

More recent books providing all necessary details frommeasure theory are (among many others);



Books on the market

John J. Benedetto and Wojciech Czaja Integration andModern Analysis Birkhauser, (2009, p.576, [1]);

or the upcoming book by Christ Heil: Introduction toHarmonic Analysis,

and books on frames, bases, or TF-analysis e.g.

C. Heil: A Basis Theory Primer. Expanded ed., ([17]);

O. Christensen: An Introduction to Frames and Riesz Bases([3]);

K. Grochenig “Foundations of TF-analysis” ([16]);

Rupert Lasser: Fourier Series and Transforms ([21]).



Fourier Analysis “in the books”

Let us take a look back into the historical development ofFOURIER ANALYSIS:First J.B. Fourier’s claim that every periodic function can beexpanded into a Fourier series was challenging the mathematicalcommunity to develop proper notions of functions and integrals(Riemann, Lebesgue) and different types of convergence. The needto properly describe measurability and exceptional sets hascertainly greatly influenced the theory of sets.The last century saw the development from classical Fourieranalysis (over Euclidean spaces Rd ) to what has been sometimecalled abstract Harmonic Analysis (over LCA groups), but alsothe invention of the theory of (tempered) distributions asa tool to extend the domain of definition to include pointmasses and polynomials.



Fourier Analysis “in the books” II

If we take a look into published books on Fourier Analysis onefinds a majority of books following the historical path:

First starting with Fourier series and Hilbert spaces

Then the Fourier transform, Riemann-Lebesgue theorem

Fourier inversion, uniqueness

Plancherel’s theorem, convolution theorem

Lp-spaces, Hausdorff-Young theorem

Poisson’s formula, Sampling, Shannon’s theorem

Fast Fourier Transforms (FFT)

perhaps some wavelets or other recent stuff

ideally indications about the extension to distributions



Fourier Analysis “in the books” III

A side aspect or rather an enabling tool in performing this task, isthe availability of suitable function spaces (in fact Banach spacesof functions or distributions) aside from the classical Lp-spaces,such as

Wiener amalgam spaces

Modulation spaces

Sobolev spaces

Besov-Triebel-Lizorkin spaces

and others ....



What are the Proposed Ingredients?

While most books still follow the historical path, starting formLebesgue integration theory, going to Lp-spaces and finish bymentioning recent developments like FFT, wavelets or frames Ithink we should reconsider the catalogue of important conceptsand how we want to introduce them.It is just like linear algebra, where the rapid development ofnumerical linear algebra is having its impact (to some extent) onthe content of modern linear algebra books, which try toaccommodate the balance between concrete and often simpleexamples, to abstract viewpoints and then not-so-trivialapplications.



The New Catalogue of Terms I

Overall I may suggest a more functional analytic approach toFourier/Harmonic Analysis. By this I mean to acknowledge thefact of life that linear spaces of signals or operators are inherentlyinfinite dimensional, so that the idea of working with a fixed basisis not a good idea, but also definitely requires to have infiniteseries, and thus completeness must play an important role.So in principle we need Banach spaces and their duals, in manycases it will be convenient to work in a Hilbert space. In the caseof dual elements we also have to allow considerations of thew∗-convergence of functionals.OVERALL: nothing exotic so far!



The New Catalogue of Terms II

Concepts and terms which are perhaps less familiar are, some ofwhich already are interesting in a linear algebra setting:

1 Banach algebras and Banach modules;

2 approximate units (e.g. Dirac sequences for convolution);

3 the four spaces associated with a linear mapping (G. Strang);

4 the SVD (singular value decomposition);

5 Banach frames and Riesz (projection) basesas extensions of generating and lin. indep. sets;

6 Banach Gelfand triples, retracts in this category;

7 unitary Banach Gelfand triple isomorphism;



A first look at the collection of space

S0Schw L1

Tempered Distr.

SO’

L2

C0

FL1



Fourier Analysis: Technical requirements I

In order to run this program properly one has to explain thenecessary terms first. The Fourier transform appears a priori to bedefined as an integral transform, namely via the equation

f (ω) =

∫Rd

f (t) · e−2πiω·t dt (3)

The inverse Fourier transform then has the form

f (t) =

∫Rd

f (ω) · e2πit·ω dω. (4)

We all know about the short-comings of the (nice and simple)Riemann integral and therefore we advocate the use ofLebesgue integration. BUT it cannot justify the inversion formula!



Fourier Analysis: Technical requirements II

There appear to be many further reasons for assuming that thetheory of Lebesgue integration learned properly is a requirement fora deeper understanding of Fourier analysis:

The Banach space(L1(Rd ), ‖ · ‖1

)is not only the “proper

domain” of the (integral) FT, but also for convolution!

f ∗ g(x) :=

∫Rd

g(x − y)f (y)dy =

∫Rd

Ty g(x)f (y)dy (5)

the convolution theorem f ∗ g = f · g .checking associativity of convolution requires Fubini;



Fourier Analysis: Technical requirements III

Still there are cases where Lebesgue integration is not be able toprovide good answers:

The Fourier inversion problems requires to introducesummability kernels and take limits of functions (in suitablefunction spaces) which are obtained by the inverse Fourierintegral;

There are discrete measures having well defined Fourier(Stieltjes) transforms, but the corresponding theory oftransformable measures is equally asymmetric!(developed by Argabright and Gil de Lamadrid)

Pure frequencies (characters) are not in the domain ofclassical Fourier transforms (their transforms should beDirac measures);

For PDE applications the Schwartz theory of tempereddistributions is the right thing anyway!



Fourier Analysis: Technical requirements IV

So what do we need in order to properly define generalizedfunctions and in particular tempered distributions.

To define them we need suitable spaces of test functions,typically infinitely differentiable functions, with compactsupport, are sufficiently rapid decay;

in the case of the Schwartz space of rapidly decreasingfunctions S(Rd ) one takes only functions which together withall their (partial) derivates decay faster than any polynomial;

in order to properly define the dual space one has to at leastindirectly put a convergence on such spaces of test functionsin order to select the continuous linear functionals;

the useful test functions spaces (such as S(Rd )) fulfillextra properties, such as Fourier invariance.



Fourier Analysis for Applied Scientists

If we look into this program and the potential role that FourierAnalysis (and more generally Harmonic Analysis) should play theteaching of Fourier Analysis we may ask ourselves:

are we teaching all of these concepts to the majority of ourmaster students (or teacher students)?

is there a chance that applied scientist such as engineers ofcomputer scientists can learn Fourier analysis in this way;

are there ways of teaching the essence of Fourier analysis in adifferent way, without making wrong claims, such as:The Dirac function is zero everywhere except atzero, but it is so immensely huge, somehowlarger than plus infinity, such that itsintegral satisfies

∫R δ0(t)dt = 1.

So should we now forget about Lebesgue’s integral??



Assessment of needs: bottom up approach

Let us try a bottom up approach, analyzing the emphasize put onthose items that take an important role in typical engineeringbooks (as opposed to the mathematical literature):

The basic terms are filters, time-invariant systems, etc.

Characterization of linear time-invariant systems asconvolution operators by means of the impulse responsefunction;

Characterization of LTIs using the transfer function

the convolution theorem connecting impulse response andtransfer function



Assessment of needs: bottom up approach: II

Asking critically what is needed.

is Lebesgue integration crucial, or even more important thanthe concept of “generalized functions”

is it OK to leave engineers with the all-important “siftingproperty” of the Dirac δ-distribution/function?

is there a way to teach distribution theory in a naturalcontext? (i.e. without topological vector spaces)

What is the role of the FFT? Just the computational littlebrother of the true Fourier transform, based on Lebesgueintegration?



Comparing with the Number Systems

Preparing for the suggestion to make use of so-called BanachGelfand Triples let us recall what the situation is when we usedifferent number system. Once we have learned to see theadvantages of the different types of numbers, taken from the fields

Q ⊂ R ⊂ C

it is easy to work with them, following properly given rules.

Inversion is only “simple” within Q!

Taking square roots (and other things) is only possible withinthe (positive part of) R;

Once C is well-defined as pairs of reals, with properlydefined addition and multiplication, the field propertiesof C follow without reflection at each step how e.g.inversion is carried out within R!



Comparing with the Number Systems, II

So in teaching analysis one has to think about the choice ofspending a lot of time on the establishment of the existence andbasic properties of the real number system, or rather assume fromthe beginning that there exists some totally ordered field,containing the rational numbers as a dense subfield, and go deeperinto analysis:Better spend time on concepts of differentiability of functions,Taylor expansions of differentiable functions, or other propertieswhich will the students enable to properly solve differentialequations or develop realistic models for application areas.Of course we are not advocating here to leave the studentswithout an awareness of what has been shown and whatis taken as granted, what could be proved in detail, andwhat is actually carried out step by step.



General Achievements in the Last Century

Recall that we have come a long way from Fourier to ModernHarmonic Analysis:

in the late 19th century concepts of pointwise convergencehad its first high period (De-la-Vallee Poussin, Dirichlet);

Set theory and integration theory have been founded, thentopology and functional analysis have been developed;

Banach algebras and in particular Gelfand’s theory provideproper foundations of Fourier analysis over LCA groups;

the existence of a Haar measure and Pontrjagin’s theorem, aswell as Plancherel’s theorem have been established;

L. Schwartz established his theory of tempereddistributions, extended by Bruhat to LCA groups;

Cooley-Tukey invent the FFT around 1965;



Assessment of needs: bottom up approach: III

Given the realistic situation that a typical, even mathematicallyoriented engineer will

never be able to follow in all details, but

nowadays students learn about linear algebra in a moreapplied spirit, using MATLABTM in their course, and cancarry out numerical experiments;

that a certain understanding of generalized functions ordistributions is unavoidable, if the computation of divergentintegrals resulting in Dirac’s or other computational tricks(which do not even have a precise meaning to thosewho have digested more theory);



Time-Frequency Analysis and Applications

Most colleagues would accept the statement that Wavelet Analysishas brought some important aspects to real life, e.g. due to itsability of bringing us a “mathematical zoom” and a good way ofcompressing image information. And clearly Fourier Analysis (inthe way it has been performed in the last century) was not sohelpful in doing this, but has Fourier Analysis been given a fairchance?Hasn’t Fourier Analysis been carrying the flavour of being import(for theoretical applications) but cumbersome and difficult when itcomes to applications. Aren’t the troubles that one has withFourier inversion evident, partially due to the fact that the buildingblocks (the pure frequency resp. complex exponentials) do notbelong to the Hilbert space

(L2(Rd ), ‖ · ‖2

), hence Fourier

Analysis and Fourier Synthesis have only a “formal meaning”?



Applications of Fourier Analysis in Modern Life

Although it is not the usual way to justify the teaching ofmathematical topics it is not unreasonable to reconsider thoseareas where Fourier Analysis and in particular the use of the FFThas a significant impact on our daily life!Let us just mention a view cases:

digital signal and image processing

mobile communication

medical imaging (e.g. tomography)

why are WAV files stored using the sampling rate 44100/sec?

how is MP3 able to compress musical information?

But what is the relationship to the Fourier Transformwe are teaching in our analysis classes?



Case Study I: Characterization of TILS

It is one of the first mathematical claims taught to electricalengineers students that every translation invariant linear system Tcan be viewed as a convolution operator. The standard rule (oftenwith some hand-waving) is the following:If one treats the case of discrete signals, say over the group Z ofintegers, then the impulse response is just the output arising if theunit-vector e0 is used as input vector. Clearly unit vectors at n ∈ Zcan be obtained from e0 by n units (left or right) thus allowing toobtain a mathematically correct characterization of TBut what kind of problems are put under the rug if we havecontinuous variables, say we are interested in TILs overthe real line R? Should/can we simply replace e0 by δ0?



Case Study I: Characterization of TILS: II

The usual way , Dirac sequences of box functions... vague andformal arguments..



Case Study I: Characterization of TILS: III

Not too long ago an obviously mathematically trained engineer,Irwin Sandberg, was observing that there is a “scandal”. Between1998 and 2004 he published a series of papers, e.g. with the title:

Continuous multidimensional systemsand the impulse response scandal.

So far the impact of his observations was quite modest, althoughhe demonstrates among others that there are bounded linearmappings from the space Cb(Rd ) (endowed with the sup-norm)into itself, which are translation invariant but cannot berepresented in the usual/expected way!It seems that the community prefers to ignore a pedanticmember rather than doing ground work on the basics!



Case Study I: Characterization of TILS: IV

But it is easy to correct the problem by replacing Cb(Rd ) byC0(Rd ), the space of continuous, complex-valued functionsvanishing at infinity (mostly because this space is separable!).Recall that by definition (justified by the Riesz-RepresentationTheorem) the dual space to

(C0(Rd ), ‖ · ‖∞

)can be called the

space of bounded (regular) Borel measures, denoted by M(Rd ).

Theorem

There is an isometric isomorphism between the Banach space ofbounded linear mappings from

(C0(Rd ), ‖ · ‖∞

)into itself which

commutes with all translation operators Tx , x ∈ Rd , and the space(M(Rd ), ‖ · ‖M) (with the norm as dual space).Given µ ∈M(Rd ) one defines the convolution operator

Cµf : f 7→ µ ∗ f (x) = µ(Tx f ), f (x) = f (−x).



Case Study I: Characterization of TILS: V

This results requires a couple of auxiliary observations, such as the(norm) density of compactly supported measures within(M(Rd ), ‖ · ‖M) in order to ensure that Cµ is not only uniformlycontinuous but in fact even in C0(Rd ), for every input signalf ∈ C0(Rd ).On the other hand all one needs is the existence of sufficientlymany functions in C0(Rd ), i.e. the density of functions withcompact support within

(C0(Rd ), ‖ · ‖∞

). This fact is easily

established using the local compactness of the additive group Rd ,and thus the proof of the theorem is valid in thecontext of general locally compact Abelian groups G !

Note that it also does not depend on the Haar measure on G .



Case Study I: Characterization of TILS: VI

There is a number of consequences from this observation, startingfrom the observation that the simple translation operators Tx

describe of course translation invariant operators on C0(G ), andcorrespond exactly to the bounded measure δx : f 7→ f (x).

Using uniform partitions of unity one can show that finite discretemeasures are dense (in the w∗-topology) in C′0(Rd ) = M(Rd ), orequivalently any TILS can be approximated in the strong operatortopology by finite linear combinations of translation operators:For ε > 0 and any finite subset F ⊂ C0 one has for a suitablesequence (an) in C and a finite sequence (xn) in Rd such that:

‖T (f )−l∑

k=1

anTxn f ‖∞ < ε ∀f ∈ F .



Case Study I: Characterization of TILS: VII

The most important consequence of the identification of TILS andbounded measures is the fact that one can transfer the Banachalgebra properties of the family of all TILS to the boundedmeasures. By definition we define composition of boundedmeasures (to be called convolution!) via the composition of thecorresponding operators. Obviously convolution is thus anassociative and bilinear operation on MbG .Since the translation operators on a LCA group commute witheach other and since they can be used to approximate generalsTILS on find out that this convolution is also commutative.Finally one shows that any µ ∈ C′0(G ) extends to all of Cb(G ) in aunique way, hence µ = µ(χs) (with χ ∈ G ) is well definedand one shows the convolution theorem: µ1 ∗ µ2 = µ1 · µ1.Again all this can be done without any integration theory!



Case Study I: Characterization of TILS: VIIa

Here let us introduce some terminology about Banach modules,which may be less familiar to the “general public”:

Definition

A Banach space (B, ‖ · ‖B) is a Banach module over a Banachalgebra (A, ·, ‖ · ‖A) if one has a bilinear mapping (a, b) 7→ a • b,from A× B into B with

‖a • b‖B ≤ ‖a‖A‖b‖B ∀ a ∈ A, b ∈ B

which behaves like an ordinary multiplication, i.e. is associative,distributive, etc.:

a1 • (a2 • b) = (a1 · a2) • b ∀a1, a2 ∈ A, b ∈ B.



Case Study I: Characterization of TILS: VIIb

Following [20] (Katznelson) we define:

Definition

A Banach space (B, ‖ · ‖B) of locally integrable functions is calleda homogeneous Banach space on Rd if it satisfies

1 ‖Tx f ‖B = ‖f ‖B ∀f ∈ B, x ∈ Rd ;

2 ‖Tx f − f ‖B → 0 for x → 0, ∀f ∈ B.

Important examples are of course the spaces B = Lp(Rd ) for1 ≤ p <∞ (which obviously are not sitting inside of L1(Rd )).It can be shown in an elementary way (just using a kind ofvector-valued Riemann-integrals) that any such space is aBanach convolution module over M(Rd ):

‖µ ∗ f ‖B ≤ ‖µ‖M‖f ‖B ∀µ ∈M(Rd ), f ∈ B. (6)



Case Study I: Characterization of TILS: VIII

In order not to play down the role of the Haar measures on anyLCA group let us remind that it is a why which enables theidentification of ordinary functions with bounded measures!In fact, one can define for k ∈ Cc(Rd ) (continuous with compactsupport) the measure

µk : h 7→∫Rd

h(x)k(x)dx , h ∈ C0(Rd ),

With some nice tricks one then has to show that

‖µk‖M =

∫Rd

|k(x)|dx ,

i.e. the Haar measure (a linear functional) applied to |k |!, andfrom there one can define L1(Rd ) as the closure of all thosemeasures in (M(Rd ), ‖ · ‖M) and and closed ideal in M(Rd ).



Summary so far

Summarizing so far we can claim that it is possible to introduce -even in the full generality of locally compact Abelian groups andwithout the use of the Haar measure basic facts about

the commutative Banach convolution algebra M(G );

the Fourier (Stieltjes) transform satisfying the convolutiontheorem;

the extension of the group action on homogeneous Banachspaces to M(G ), which is thus taking the role of the groupalgebra over finite Abelian groups.

the technical tool to achieve this are fine partitions of unity!

This also shows that one should not let oneself be distractedfrom certain technical questions which arise from measuretheoretical considerations!



Recalling some Basic Questions

Most often the “naturalness” of the assumption of Lebesgueintegrability, i.e. the restriction to L1(Rd ) is coming fromconsiderations concerning the existence (!!??!!) of the Fouriertransform or some convolution. But WHAT DO WE MEAN, whenmaking one of the following claims?

The function has a Fourier transform?

The convolution f ∗ g of two function f , g exists?

The convolution operator f 7→ f ∗ g is well defined on a givenBanach space of functions

(B, ‖ · ‖B

);

what is the validity range for the convolution theorem

σ1 ∗ σ2 = σ1 ∗ σ2



Recalling some Basic Questions, II

Let us consider the first question. We have many options

The Fourier transform is well defined pointwise as a Lebesgueintegral. This is of course just a slightly reformulation of theassumption f ∈ L1(Rd )! (and thus not very interesting);

But isn’t then the Fourier transform of the SINC function:SINC (t) = sin(πt)/πt, t ∈ R \ {0} not well defined, althoughone could take the Fourier integral in the improperRiemannian sense. Should we prohibit this option, but how farcan we go in making exceptions?

certainly SINC has an (inverse) Fourier transform (thebox-function) in the L2-sense.

in such situations: should we allow a combinationof summability methods, but require pointwise convergence?



Recalling some Basic Questions, III

What about the well-definedness of linear operators for bilinearoperations such as convolution? When deserves a function (ordistribution) to be called a convolution kernel for a Banach space(B, ‖ · ‖B

)(such as L2(Rd ) or Lp(Rd ))?

Again there are various options:a) the strict one: For every h ∈ B the Lebesgue integral∫Rd h(x − y)f (y)dy exists almost everywhere and defines a new

element to be called f ∗ h ∈ B;b) The integral is well-defined for test functions h ∈ B, but

‖f ∗ h‖B ≤ Cf ‖h‖B

for all such h, assuming that they are dense in (B, ‖ · ‖B);



Recalling some Basic Questions, IIIb

In many situations there will be now difference, e.g. if bothf , g ∈ L2(Rd ) then the Cauchy-Schwarz inequality guarantees theexistence of convolution integrals and

|f ∗ g(x)| ≤ ‖f ‖2‖g‖2 ∀f , g ∈ L2

But what about functions g /∈ L2(Rd ) which happen to have abounded Fourier transform, and hence define a bounded Fouriermultiplier thanks to Plancherel’s Fourier transform. Chirpfunctions of the form t 7→ e iαt2

are typical examples, because theirFT is another chirp.In other words, the pointwise concept is imposing restrictionswhich are not suitable from a functional analytic viewpoint.This was also the cause of troubles in work on temperedLp-functions by K. McKennon in the 70th ([12]).



The Proper Setting from a TF-viewpoint

Surprisingly many classical questions can be given a more elegantinterpretation using function spaces which would be mostlyconsidered as suitable and relevant for TF applications (only), suchas the Segal algebra S0(Rd ) (resp. S0(G )) and its dual.Originally ([6]) it got its symbol because it is the smallest(therefore the index zero!) in a class of so-called Segal algebras asstudied by Hans Reiter ([23]).It is defined as follows: A bounded, continuous and integrablefunction f belongs S0(Rd ) if (and only if) it has a (Riemann-)integrable Short-time Fourier transform Vg0f , defined as

‖f ‖S0 := ‖Vg0f ‖L1(Rd×Rd )

The assumption is so strong that the same space arises evenif one just assumes a priori that f is a tempered distribution.



The Proper Setting from a TF-viewpoint

It is not difficult to show that the space is the same if g0 is replacedby any other non-zero function from the Schwartz space, or evenfrom S0(Rd ) itself. In such a way the space can be defined forgeneral LCA groups (using e.g. smooth and compactly supportedfunctions instead of g0, or a some g ∈ L1(G ) with compactlysupported Fourier transform, i.e. a band-limited L1-function).The advantage of the choice g = g0 is the Fourier invariance of g0,which in turn implies that S0(Rd ) is isometrically invariant underthe Fourier transform!Depending on the viewpoint, one can say that it is one of the fewFourier invariant Banach spaces (aside from L2(Rd )), sitting insideof L2(Rd ), or otherwise, that the Fourier transform is not onlyisometric on S0(Rd ), but also isometric in the L2-norm andtherefore extends to a unitary transformation for L2(Rd ),viewed as the completion of S0(Rd ) with respect to ‖ · ‖2.



Generalized Stochastic Processes

Definition

1 Let H be an arbitrary Hilbert space. A bounded linearmapping ρ : S0(G )→ H is called a generalized stochasticprocess (GSP).

2 A GSP ρ is called (wide sense time-) stationary, if

(ρ(f )|ρ(g)) = (ρ(Tx f )|ρ(Tx g)) ∀ x ∈ G and ∀ f , g ∈ S0(G ).

3 A GSP ρ is called (wide sense) frequency stationary, if

(ρ(f )|ρ(g)) = (ρ(Mt f )|ρ(Mtg)) ∀ t ∈ G and ∀ f , g ∈ S0(G ).

4 A time- and frequency-stationary GSP is called white noise.



Generalized Stochastic Processes II

Definition

1 A GSP ρ is called bounded, if ∃ c > 0 such that

‖ρ(f )‖H ≤ c‖f ‖∞ ∀ f ∈ S0(G ).

2 A GSP ρ is called variation-bounded (V-bounded), if∃ c > 0 such that

‖ρ(f )‖H ≤ c‖f ‖∞ ∀ f ∈ S0(G ).

3 A GSP ρ is called orthogonally scattered if

supp(f ) ∩ supp(g) = ∅ =⇒ ρ(f ) ⊥ ρ(g) for f , g ∈ S0(G ).



Generalized Stochastic Processes III

Due to the tensor product property of S0 :

S0(G )⊗S0(G ) = S0(G × G )

(cf. [6] Theorem 7 D) it is justified to hope that the followingdefinition determines an element of S0(G × G ).

Definition 6: Let ρ be a GSP. The autocovariance (orauto-correlation) distribution σρ is defined as:〈σρ, f ⊗ g〉 := (ρ(f )|ρ(g)) ∀ f , g ∈ S0(G ).



Generalized Stochastic Processes IV

Theorem 1: For a GSP ρ the following properties are equivalent:a) ρ stationary ⇐⇒ σρ diagonally invariant, i.e.

L(x ,x)σρ = σρ ∀ x ∈ G ;

b) ρ bounded ⇐⇒ σρextends in a unique way to a bimeasure on G × G ;c) ρ orthogonally scattered⇐⇒ σρ is supported by the diagonal, i.e.supp(σρ) ⊆ ∆G := {(x , x) | x ∈ G} ;⇐⇒ ∃ positive and translation bounded measure τρ with:

〈σρ, f ⊗ g〉 = 〈τρ, fg〉 ∀ f , g ∈ S0(G ).



Further applications areas

There is a large variety of topics, where the same setting appearsto be most appropriate, mostly because it is technically easier thanthe use of Schwartz-Bruhat functions, and on the other handmore general (S(G ) ⊂ S0(G )), in particular work onnon-commutative geometry by Connes and Rieffel (via the work ofFranz Luef);



Realizable constructive approaches vs. constructive

We will illustrate this by the following:Let us look at the spline-type spaces (e.g. cubic splines on R) inLp-spaces, for p 6= 2. Then this spaces are closed subspaces of(Lp(R), ‖ · ‖p

), which is uniformly convex, hence there is a best

approximation to any given f ∈ Lp(R) by an element from thecorresponding spline-type space. But this is not a constructiveapproach.I contrast we have quite explicit methods (so-called constructiveapproximation) to obtain the (orthogonal) approximation in thecase p = 2, and the corresponding operator can be shown to bealso (uniformly) bounded on the family of Lp-spaces, for the fullrange of 1 ≤ p ≤ ∞.See [8] and an upcoming Springer BRIEF for discussions ofthe problems arising with this, for the multi-window case.



Main aspect: Factorize through finite dimensions

The general idea - at least in the context of Gabor analysis - ofthis refined approach (merging concepts of functional analysis andnumerical analysis) is to approximate a continuous problem by thecorresponding problem in the finite context, i.e. on a suitablychoosen finite Abelian group, with the need to describe also theprocess of returning from discrete/finite data to the continuousdomain. Usually this is done by quasi-interpolation. The mostimportant special case of quasi-interpolation is piecewise linearinterpolation, (quasi-interpolation using linear splines), while theoperator Q(d) =

∑dλTλϕ is not interpolating anymore (e.g. for

d = (f (λ))) for the case of cubic B-splines.



MATLAB illustration of the situation

−300 −200 −100 0 100 200 300

−0.05

0

0.05

−300 −200 −100 0 100 200 300

−0.05

0

0.05

red: sampling values of cubid quasi−interpolator

−300 −200 −100 0 100 200 300

−0.05

0

0.05

piecewise linear interpolation



Connections to related situations

Relevant results (mostly due to Norbert Kaiblinger and hgfei)

quasi-interpolation in the Fourier algebra (S0-setting);

in particular: realizing the (continuous) FT using FFTmethods

robustness results (varying the lattice constants, etc.),required to find a suitable finite model (!periodicityassumptions);



Connections to related situations

Wavelet multipliers (shearlet multipliers, Toeplitz operatorsetc.);

Quality of operators depending only on the rough propertiesof the symbol, at least for “nice” windows;

Connection between discrete variants of these operators andcorresponding continuous frame multipliers;



END of adjusted material



Goals of this lecture

show that the usual generalizations of linear algebra conceptsto the Hilbert space case (namely linear independence andtotality) are inappropriate in many cases;

that frames and Riesz bases (for subspaces) are the rightgeneralization to Hilbert spaces;

that Hilbert spaces are themselves a too narrow concept andshould be replaced Banach Gelfand Triples, ideally isomorphicto the canonical ones (`1, `2, `∞);

Describing the situation of frames or Riesz bases via commu-tative diagrams allows to extend this notation to BGTs;

Demonstrate by examples (Fourier transform, kerneltheorem) that this viewpoint brings us very close to thefinite-dimensional setting!



OVERVIEW over this lecture 60- MINUTES

This is a talk about the frames viewing frames as RETRACTS,i.e. a construction which can be done in any category;

about the ubiquity of Banach Gelfand Triples ;

provides a setting very similar to the finite dimensional setting;

showing how easy they are to use;

showing some applications in Fourier Analysis;

indicating its relevance for numerical applications;

and for teaching purposes;

that it is a good vehicle to transfer algebraic facts(over finite Abelian group to the setting of LCA groups);

perhaps change your view on Fourier Analysis.



General Considerations

One of the specifically interesting aspects of Gabor analysis is thefact, that it can be realized over general LCA groups.Typically we discuss mathematical questions on Gabor Analysis inthe context of Euclidean spaces, i.e. for R or Rd , while theimplementations take place for finite, discrete signals, modeled asfunction on the cyclic groups, e.g. over ZN (cyclic group of orderN) or in the case of image processing over ZM × ZN .In this setting one can much better see what the linear algebrabackground is (kernels, linear dependence) and the algebraicproperties, while in the general case (much) more functionalanalysis comes in (a variety of norms, types of convergence, etc.).



ANALYSIS: Calculating with all kind of numbers

We teach in our courses that there is a huge variety of NUMBERS,but for our daily life rationals, reals and complex numbers suffice.The most beautiful equation

e2πi = 1.

It uses the exponential function, with a (purely) imaginaryexponent to get a nice result, more appealing than (the equivalent)

cos(2π) + i ∗ sin(2π) = 1 in C.

But actual computation are done for rational numbers only!! Recall

Q ⊂ R ⊂ C



ANALYSIS: Spaces used to describe the Fourier Transform

S0Schw

Tempered Distr.Ultradistr.

SO’L2

S0SchwFL1

Tempered Distr.

SO’L2

C0

L1



Hausdorff Young Theorem for the Fourier Transform

FLp(Rd ) ⊆ Lq(Rd ), 1 ≤ p ≤ 2,1

q+

1

p= 1.



Wiener Amalgam spaces, Wiener Algebra, etc.



Frames in Hilbert Spaces: Classical Approach

Definition

A family (fi )i∈I in a Hilbert space H is called a frame if there existconstants A,B > 0 such that for all f ∈ H

A‖f ‖2 ≤∑i∈I

|〈f , fi 〉|2 ≤ B‖f ‖2 (7)

It is well known that condition (7) is satisfied if and only if theso-called frame operator is invertible, which is given by

Definition

S(f ) :=∑i∈I

〈f , fi 〉fi , for f ∈ H,



Frames in Hilbert Spaces: Classical Approach II

The obvious fact S ◦ S−1 = Id = S−1 ◦ S implies that the(canonical) dual frame (fi )i∈I , defined by fi := S−1(fi ) has theproperty that one has for f ∈ H:

Definition

f =∑i∈I

〈f , fi 〉fi =∑i∈I

〈f , fi 〉fi (8)

Moreover, applying S−1 to this equation one finds that the family(fi )i∈I is in fact a frame, whose frame operator is just S−1, andconsequently the “second dual frame” is just the original one.



Frames in Hilbert Spaces: Approach III

Since S is positive definite in this case we can also get to a moresymmetric expression by defining hi = S−1/2fi . In this case one has

f =∑i∈I

〈f , hi 〉hi for all f ∈ H. (9)

The family (hi )i∈I defined in this way is called the canonical tightframe associated to the given family (gi )i∈I . It is in some sense theclosest tight frame to the given family (fi )i∈I .



Where did frames come up? Historical views:

I think there is a historical reason for frames to pop up in thesetting of separable Hilbert spaces H. The first and fundamentalpaper was by Duffin and Schaeffer ([5]) which gained popularity inthe “painless” paper by Daubechies, Grossmann and Y. Meyer([4]). It gives explicit constructions of tight Wavelet as well asGabor frames. For the wavelet case such dual pairs are are alsoknown due to the work of Frazier-Jawerth, see [9, 10]. Suchcharacterizations (e.g. via atomic decompositions, with control ofthe coefficients) can in fact seen as prerunners of the concept ofBanach frames to be discussed below.These methods are closely related to the Fourier description offunction spaces (going back to H. Triebel and J. Peetre) usingdyadic partitions of unity on the Fourier transform side.



Dyadic Partitions of Unity and Besov spaces

0 1 2 3 4 5 6 7 8−0.2

0

0.2

0.4

0.6

0.8

1

20 40 60 80 100 120 140 160 180

0

0.2

0.4

0.6

0.8

1

dyadic BUPU by transfer from uniform BUPU on R



Where did frames come up? Historical views II

The construction of orthonormal wavelets (in particular the firstconstructions by Y. Meyer and Lemarie, and subsequently thefamous papers by Ingrid Daubechies), with prescribed degree ofsmoothness and even compact support makes a big difference tothe Gabor case.In fact, the Balian-Low theorem prohibits the existence of (Riesz-or) orthogonal Gabor bases with well TF-localized atoms, henceone has to be content with Gabor frames (for signal expansions) orGabor Riesz basic sequences (for mobile communication such asOFDM).This also brings up a connection to filter banks, which in the caseof Gabor frames has been studied extensively by H. Bolcskeiand coauthors ( see [2]).



LINEAR ALGEBRA: Gilbert Strang’s FOUR SPACES

Let us recall the standard linear algebra situation. Given somem × n -matrix A we view it as a collection of column resp. as acollection of row vectors. We have:

row-rank(A) = column-rank(A)Each homogeneous linear system of equations can be expressed inthe form of scalar products1 we find that

Null(A) = Rowspace(A)⊥

and of course (by reasons of symmetry) for A′ := conj(At):

Null(A′) = Colspace(A)⊥

1Think of 3x + 4y + 5z = 0 is just another way to say that the vectorx = [x , y , z] satisfies 〈x, [3, 4, 5]〉 = 0.



Geometric interpretation of matrix multiplication

Since clearly the restriction of the linear mapping x 7→ A ∗ x

Rn

Row(A) Col(A) ⊆ Rm-T = T|row(A)

R

?

PRow

@@@@@@@@R

T : x → A ∗ x



Geometric interpretation of matrix multiplication

Null(A) ⊆ Rn

Row(A) Col(A) ⊆ Rm-

T = T|row(A)

inv(T )

?

PRow

@@@@@@@@@@R

T T ′

Rm ⊇ Null(A′)

?

�

PCol

��

��

��

?

T = T ◦ PRow , pinv(T ) = inv(T ) ◦ PCol .



Four spaces and the SVD

The SVD (the so-called Singular Value Decomposition) of amatrix, described in the MATLAB helpful as a way to write A as

A = U ∗ S ∗ V ′

, where the columns of U form an ON-Basis in Rm and the columnsof V form an ON-basis for Rn, and S is a (rectangular) diagonalmatrix containing the non-negative singular values (σk ) of A. Wehave σ1 ≥ σ2 . . . σr > 0, for r = rank(A), while σs = 0 for s > r .In standard description we have for A and pinv(A) = A+:

A ∗ x =r∑

k=1

σk〈x , vk〉uk , A+ ∗ y =r∑

k=1

1

σk〈y , uk〉vk .



Generally known facts in this situation

The Four Spaces are well known from LINEAR ALGEBRA, e.g. inthe dimension formulas:

ROW-Rank of A equals COLUMN-Rank of A.

The defect (i.e. the dimension of the Null-space of A) plus thedimension of the range space of A (i.e. the column space of A)equals the dimension of the domain space Rn. Or in terms oflinear, homogeneous equations: The dimension of set of allsolution to the homogeneous linear equations equals the number ofvariables minus the dimension of the column space of A.The SVD also shows, that the isomorphism between theRow-space and the Column-space can be described by a diagonalmatrix, if suitable orthonormal basis for these spaces are used.



Consequences of the SVD

We can describe the quality of the isomorphism T by looking at itscondition number, which is σ1/σr , the so-called Kato-conditionnumber of T .It is not surprising that for normal matrices with A′ ∗ A = A ∗ A′

one can even have diagonalization, i.e. one can choose U = V ,because

Null(A) =always Null(A′ ∗ A) = Null(A ∗ A′) = Null(A′).

The most interesting cases appear if a matrix has maximal rank,i.e. if rank(A) = min(m, n), or equivalently if one of the twoNull-spaces is trivial. Then we have either linear independentcolumns of A (injectivity of T >> RIESZ BASIS forsubspaces) or the columns of A span all of Rm

( i.e. Null(A′) = {0}): FRAME SETTING!



Geometric interpretation: linear independent set > R.B.

Row(A) = Rn Col(A) ⊆ Rm-

T = T|row(A)

inv(T ) = pinv(A)

T ′

Rm ⊇ Null(A′)

?

�

PCol

��

��

��

?



Geometric interpretation: generating set > FRAME

Null(A) ⊆ Rn

Row(A) Col(A) = Rm-

T = T|row(A)

inv(T ) = A′

?

PRow

@@@@@@@@@@R

T

�



The frame diagram for Hilbert spaces:

If we consider A as a collection of column vectors, then the role ofA′ is that of a coefficient mapping: f 7→ (〈f , fi 〉).

`2(I )

H C(H)-C

� R ?

P

��

��

��

R

This diagram is fully equivalent to the frame inequalities (6).



Riesz basic sequences in Hilbert spaces:

The diagram for a Riesz basis (for a subspace), nowadays called aRiesz basic sequence looks quite the same.In fact, from an abstract sequence there is no! difference, just likethere is no difference (from an abstract viewpoint) between amatrix A and the transpose matrix A′.However, it makes a lot of sense to think that in one case thecollection of vectors (making up a Riesz BS) spans the (closed)subspace of H by just taking all the infinite linear combinations(series) with `2-coefficients.In this way the synthesis mapping c 7→

∑i ci gi from `2(I ) into the

closed linear span is surjective, while in the frame case theanalysis mapping f 7→ (〈f , gi 〉) from H into `2(I ) is injective(with bounded inverse).



Frames versus atomic decompositions

Although the definition of frames in Hilbert spaces emphasizes theaspect, that the frame elements define (via the Rieszrepresentation theorem) an injective analysis mapping, theusefulness of frame theory rather comes from the fact that framesallow for atomic decompositions of arbitrary elements f ∈ H.One could even replace the lower frame bound inequality in thedefinition of frames by assuming that one has a Bessel sequence(i.e. that the upper frame bound is valid) with the property thatthe synthesis mapping from `2(I ) into H, given by c 7→

∑i ci gi is

surjective onto all of H.Analogously one can find Riesz bases interesting (just like linearindependent sets) because they allow to uniquely determinethe coefficients of f in their closed linear span on that closedsubspace of H.



A hierarchy of conditions 1

While the following conditions are equivalent in the case of a finitedimensional vector space (we discuss the frame-like situation) onehas to put more assumptions in the case of separable Hilbertspaces and even more in the case of Banach spaces.Note that one has in the case of an infinite-dimensional Hilbertspace: A set of vectors (fi )i∈I is total in H if and only if theanalysis mapping f 7→ (〈f , gi 〉) is injective. In contrast to theframe condition nothing is said about a series expansion, and infact for better approximation of f ∈ H a completely different finitelinear combination of g ′i s can be used, without any control on the`2-norm of the corresponding coefficients.THEREFORE one has to make the assumption that the rangeof the coefficient mapping has to be a closed subspace of`2(I ) in the discussion of frames in Hilbert spaces.




In the case of Banach spaces one even has to go one step further.Taking the norm equivalence between some Banach space normand a corresponding sequence space norm in a suitable Banachspace of sequences over the index set I (replacing `2(I ) for theHilbert space) is not enough!In fact, making such a definition would come back to theassumption that the coefficient mapping C : f 7→ (〈f , gi 〉) allows toidentify with some closed subspace of that Banach space ofsequences. Although in principle this might be a useful concept itwould not cover typical operations, such as taking Gaborcoefficients and applying localization or thresholding, as themodified sequence is then typically not in the range of thesampled STFT, but resynthesis should work!




What one really needs in order to have the diagram is theidentification of the Banach space under consideration (modulationspace, or Besov-Triebel-Lozirkin space in the case of waveletframes) with a close and complemented subspace of a larger spaceof sequences (taking the abstract position of `2(I ).To assume the existence of a left inverse to the coefficient mappingallows to establish this fact in a natural way. Assume that R is theleft inverse to C. Then C ◦ R is providing the projection operator(the orthogonal projection in the case of `2(I ), if the canonicaldual frame is used for synthesis) onto the range of C. The converseis an easy exercise: starting from a projection followed by theinverse on the range one obtains a right inverse operator R.




The above situation (assuming the validity of a diagram and theexistence of the reconstruction mapping) is part of the definition ofBanach frames as given by K. Grochenig in [13].Having the classical situation in mind, and the spirit of frames inthe Hilbert spaces case one should however add two moreconditions:In order to avoid trivial examples of Banach frames one shouldassume that the associated Banach space

(B, ‖ · ‖B

)of sequences

should be assumed to be solid, i.e. satisfy that |ai | ≤ |bi | for alli ∈ I and b ∈ B implies a ∈ B and ‖a‖B ≤ ‖b‖B.Then one could identify the reconstruction mapping R with thecollection of images of unit vectors hi := R(~ei ), where ~ei is theunit vector at i ∈ I . Moreover, unconditional convergenceof a series of the form

∑i ci hi would be automatic.




Instead of going into this detail (including potentially thesuggestion to talk about unconditional Banach frames) I wouldlike to emphasize another aspect of the theory of Banach frames.According to my personal opinion it is not very interesting todiscuss individual Banach frames, or the existence of some Banachframes with respect to some abstract Banach space of sequences,even if the above additional criteria apply.The interesting cases concern situations, where the coefficient andsynthesis mapping concern a whole family of related Banachspaces, the setting of Banach Gelfand triples being the minimal(and most natural) instance of such a situation.A comparison: As the family, consisting of father, mother and thechild is the foundation of our social system, Banach GelfandTriples are the prototype of families, sometimes scales of Banachspaces, the “child” being of course our beloved Hilbert space.



Banach Gelfand Triples and Rigged Hilbert space

The next term to be introduced are Banach Gelfand Triples.There exists already and established terminology concerning triplesof spaces, such as the Schwartz triple consisting of the spaces(S,L2,S ′)(Rd ), or triples of weighted Hilbert spaces, such as(L2

w ,L2,L2

1/w ), where w(t) = (1 + |t|2)s/2 for some s > 0, which is- via the Fourier transform isomorphic to another (“Hilbertian”)Gelfand Triple of the form (Hs ,L

2,Hs′), with a Sobolev space and

its dual space being used e.g. in order to describe the behaviour ofelliptic partial differential operators.The point to be made is that suitable Banach spaces, in factimitating the prototypical Banach Gelfand triple (`1, `2, `∞)allows to obtain a surprisingly large number of resultsresembling the finite dimensional situation.



A Classical Example related to Fourier Series

There is a well known and classical example related to the moregeneral setting I want to describe, which - as so many things - goback to N. Wiener. He introduced (within L2(U)) the space(A(U), ‖ · ‖A

)of absolutely convergent Fourier series. Of course

this space sits inside of(L2(U), ‖ · ‖2

)as a dense subspace, with

the norm ‖f ‖A :=∑

n∈Z |f (n)|.Later on the discussion about Fourier series and generalizedfunctions led (as I believe naturally) to the concept ofpseudo-measures, which are either the elements of the dual of(A(U), ‖ · ‖A

), or the (generalized) inverse Fourier transforms of

bounded sequences, i.e. F−1(`∞(Z)).In other words, this extended view on the Fourier analysis operatorC : f 7→ (f (n)n∈Z) on the BGT (A,L2,PM) into (`1, `2, `∞)is the prototype of what we will call a BGT-isomorphism.



The visualization of a Banach Gelfand Triple

The S0 Gelfand triple

S0

S0’

L2



Rethinking shortly the Fourier Transform

Since the Fourier transform is one of the central transforms, bothfor abstract harmonic analysis, engineering applications andpseudo-differential operators let us take a look at it first. People(and books) approach it in different ways and flavours:

It is defined as integral transform (Lebesgue!?);

It is computed using the FFT (what is the connection);

Should engineers learn about tempered distributions?

How can we reconcile mathematical rigor and still stay intouch with applied people (physics, engineering).



The finite Fourier transform (and FFT)

For practical applications the discrete (finite) Fourier transform isof upmost importance, because of its algebraic properties [jointdiagonalization of circulant matrices, hence fast multiplication ofpolynomials, etc.] and its computational efficiency(FFT algorithms of signals of length N run in Nlog(N) time, forN = 2k , due to recursive arguments).It maps a vector of length n onto the values of the polynomialgenerated by this set of coefficients, over the unit roots of order non the unit circle (hence it is a Vandermonde matrix). It is aunitary matrix (up to the factor 1/

√n) and maps pure frequencies

onto unit vectors (engineers talk of energy preservation).



The Fourier Integral and Inversion

If we define the Fourier transform for functions on Rd using anintegral transform, then it is useful to assume that f ∈ L1(Rd ), i.e.that f belongs to the space of Lebesgues integrable functions.

f (ω) =

∫Rd

f (t) · e−2πiω·t dt (10)

The inverse Fourier transform then has the form

f (t) =

∫Rd

f (ω) · e2πit·ω dω, (11)

Strictly speaking this inversion formula only makes sense under theadditional hypothesis that f ∈ L1(Rd ). One often speaks of Fourieranalysis being the first step, and the Fourier inversion as a methodto build f from the pure frequencies (we talk of Fourier synthesis).



The classical situation with Fourier

Unfortunately the Fourier transform does not behave well withrespect to L1, and a lot of functional analysis went into fightingthe problems (or should we say symptoms?)

1 For f ∈ L1(Rd ) we have f ∈ C0(Rd ) (but not conversely, norcan we guarantee f ∈ L1(Rd ));

2 The Fourier transform f on L1(Rd ) ∩ L2(Rd ) is isometric inthe L2-sense, but the Fourier integral cannot be writtenanymore;

3 Convolution and pointwise multiplication correspond to eachother, but sometimes the convolution may have to be taken asimproper integral, or using summability methods;

4 Lp-spaces have traditionally a high reputation amongfunction spaces, but tell us little about f .



Effects of Sampling and Periodization: Poisson’s formula

−500 0 500

−0.05

0

0.05

signals

−500 0 500

−2

0

2

spectra

−500 0 500

−0.05

0

0.05

−500 0 500

−0.5

0

0.5

−500 0 500

−0.05

0

0.05

−500 0 500

−10

−5

0

5

10



A schematic description of the situation

L1

L2

C0

FL1

the classical Fourier situation



The way out: Test Functions and Generalized Functions

The usual way out of this problem zone is to introduce generalizedfunctions. In order to do so one has to introduce test functions,and give them a reasonable topology (family of seminorms), sothat it makes sense to separate the continuous linear functionalsfrom the pathological ones. The “good ones” are admitted andcalled generalized functions, since most reasonable ordinaryfunctions can be identified (uniquely) with a generalized function(much as 5/7 is a complex number!).If one wants to have Fourier invariance of the space ofdistributions, one must Fourier invariance of the space of testfunctions (such as S(Rd )). If one wants to have - in addition -also closedness with respect to differentiation one has to takemore or less S(Rd ). But there are easier alternatives.



A schematic description of the situation

S0Schw L1

Tempered Distr.

SO’

L2

C0

FL1



The Banach space(S0(Rd ), ‖ · ‖S0

)Without differentiability there is a minimal, Fourier andisometrically translation invariant Banach space (called(S0(Rd ), ‖ · ‖S0

)or (M1(Rd ), ‖ · ‖M1)), which will serve our

purpose. Its dual space (S0′(Rd ), ‖ · ‖S0

′) is correspondingly thelargest among all Fourier invariant and isometrically translationinvariant “objects” (in fact so-called local pseudo-measures orquasimeasures, orginally introduced in order to describe translationinvariant systems as convolution operators).Although there is a rich zoo of Banach spaces around (one canchoose such a family, the so-called Shubin classes - to intersect inthe Schwartz class and their union is corresondingly S ′(Rd )), wewill restrict ourselves to Banach Gelfand Triples, mostlyrelated to (S0,L

2,S0′)(Rd ).



repeated: SOGELFTR

The S0 Gelfand triple

S0

S0’

L2



The key-players for time-frequency analysis

Time-shifts and Frequency shifts

Tx f (t) = f (t − x)

and x , ω, t ∈ Rd

Mωf (t) = e2πiω·t f (t) .

Behavior under Fourier transform

(Tx f ) = M−x f (Mωf ) = Tω f

The Short-Time Fourier Transform

Vg f (λ) = 〈f ,MωTtg〉 = 〈f , π(λ)g〉 = 〈f , gλ〉, λ = (t, ω);



A Typical Musical STFT



A Banach Space of Test Functions (Fei 1979)

A function in f ∈ L2(Rd ) is in the subspace S0(Rd ) if for somenon-zero g (called the “window”) in the Schwartz space S(Rd )

‖f ‖S0 := ‖Vg f ‖L1 =

∫∫Rd×Rd

|Vg f (x , ω)|dxdω <∞.

The space(S0(Rd ), ‖ · ‖S0

)is a Banach space, for any fixed,

non-zero g ∈ S0(Rd )), and different windows g define the samespace and equivalent norms. Since S0(Rd ) contains the Schwartzspace S(Rd ), any Schwartz function is suitable, but alsocompactly supported functions having an integrable Fouriertransform (such as a trapezoidal or triangular function) aresuitable. It is convenient to use the Gaussian as a window.



Basic properties of M1 = S0(Rd )

Lemma

Let f ∈ S0(Rd ), then the following holds:

(1) π(u, η)f ∈ S0(Rd ) for (u, η) ∈ Rd × Rd , and‖π(u, η)f ‖S0 = ‖f ‖S0 .

(2) f ∈ S0(Rd ), and ‖f ‖S0 = ‖f ‖S0 .

In fact,(S0(Rd ), ‖ · ‖S0

)is the smallest non-trivial Banach space

with this property, and therefore contained in any of the Lp-spaces(and their Fourier images).There are many other independent characterization of this space,spread out in the literature since 1980, e.g. atomic decompo-sitions using `1-coefficients, or as W(FL1, `1) = M0

1,1(Rd ).



Basic properties of M∞(Rd ) = S0′(Rd )

It is probably no surprise to learn that the dual space of(S0(Rd ), ‖ · ‖S0

), i.e. S0

′(Rd ) is the largest (reasonable) Banachspace of distributions (in fact local pseudo-measures) which isisometrically invariant under time-frequency shiftsπ(λ), λ ∈ Rd × Rd .As an amalgam space one hasS0′(Rd ) = W(FL1, `1)

′= W(FL∞, `∞)(Rd ), the space of

translation bounded quasi-measures, however it is much better tothink of it as the modulation space M∞(Rd ), i.e. the space of alltempered distributions on Rd with bounded Short-time Fouriertransform (for an arbitrary 0 6= g ∈ S0(Rd )).Consequently norm convergence in S0

′(Rd ) is just uniformconvergence of the STFT, while certain atomic characterizations of(S0(Rd ), ‖ · ‖S0

)imply that w∗-convergence is in fact equivalent to

locally uniform convergence of the STFT. – Hifi recordings!



BANACH GELFAND TRIPLES: a new category

Definition

A triple, consisting of a Banach space B, which is dense in someHilbert space H, which in turn is contained in B′ is called aBanach Gelfand triple.

Definition

If (B1,H1,B′1) and (B2,H2,B

′2) are Gelfand triples then a linear

operator T is called a [unitary] Gelfand triple isomorphism if

1 A is an isomorphism between B1 and B2.

2 A is [a unitary operator resp.] an isomorphism between H1

and H2.

3 A extends to a weak∗ isomorphism as well as a norm-to-normcontinuous isomorphism between B′1 and B′2.



Banach Gelfand Triples, ctc.

In principle every CONB (= complete orthonormal basis)Ψ = (ψi )i∈I for a given Hilbert space H can be used to establishsuch a unitary isomorphism, by choosing as B the space ofelements within H which have an absolutely convergent expansion,i.e. satisfy

∑i∈I |〈x , ψi 〉| <∞.

For the case of the Fourier system as CONB for H = L2([0, 1]), i.e.the corresponding definition is already around since the times ofN. Wiener: A(U), the space of absolutely continuous Fourier series.It is also not surprising in retrospect to see that the dual spacePM(U) = A(U)′ is space of pseudo-measures. One can extend theclassical Fourier transform to this space, and in fact interpret thisextended mapping, in conjunction with the classical Planchereltheorem as the first unitary Banach Gelfand triple isomorphism,between (A,L2,PM)(U) and (`1, `2, `∞)(Z).



The BGT (S0,L2,S0

′) and Wilson Bases

Among the many different orthonormal bases the wavelet basesturn out to be exactly the ones which are well suited tocharacterize the distributions by their membership in the classicalBesov-Triebel-Lizorkin spaces.For the analogue situation (using the modulation operator insteadof the dilation, resp. the Heisenberg group instead of the“ax+b”-group) one finds that local Fourier bases resp. theso-called Wilson-bases are the right tool. They are formed fromtight Gabor frames of redundancy 2 by a particular way ofcombining complex exponential functions (using Euler’s formula) tocos and sin functions in order to build a Wilson ONB for L2(Rd ).In this way another BGT-isomorphism between (S0,L

2,S0′)

and (`1, `2, `∞) is given, for each concrete Wilson basis.



The Fourier transform as BGT automorphism

The Fourier transform F on Rd has the following properties:

1 F is an isomorphism from S0(Rd ) to S0(Rd ),

2 F is a unitary map between L2(Rd ) and L2(Rd ),

3 F is a weak* (and norm-to-norm) continuous bijection fromS0′(Rd ) onto S0

′(Rd ).

Furthermore, we have that Parseval’s formula

〈f , g〉 = 〈f , g〉 (12)

is valid for (f , g) ∈ S0(Rd )× S0′(Rd ), and therefore on each level

of the Gelfand triple (S0,L2,S0

′)(Rd ).



The w ∗− topology: a natural alternative

It is not difficult to show, that the norms of (S0,L2,S0

′)(Rd )correspond to norm convergence in (L1,L2,L∞)(R2d ).The FOURIER transform, viewed as a BGT-automorphism isuniquely determined by the fact that it maps pure frequencies ontothe corresponding point measures δω.This is a typical case, where we can see, that the w∗-continuityplays a role, and where the fact that δx ∈ S0

′(Rd ) as well asχs ∈ S0

′(Rd ) are important.In the STFT-domain the w∗-convergence has a particular meaning:a sequence σn is w∗-convergent to σ0 if Vg (σn)(λ)→ Vg (σ0)(λ)uniformly over compact subsets of the TF-plane (for one orany non-zero g ∈ S0(Rd )).



Classical Results in Banach Space Theory

Wiener’s inversion theorem:

Theorem

Assume that h ∈ A(U) is free of zeros, i.e. that h(t) 6= 0 for allt ∈ U. Then the function g(t) := 1/h(t) belongs to A(U) as well.

The proof of this theorem is one of the nice applications of aspectral calculus with methods from Banach algebra theory.This result can be reinterpreted in our context as a results whichstates:Assume that the pointwise multiplication operator f 7→ h · f isinvertible as an operator on

(L2(U), ‖ · ‖2

), and also a

BGT-morphism on (A,L2,PM) (equivalent to the assumptionh ∈ A(U)!), then it is also continuously invertible asBGT-morphism.



SOGTr-results in Banach Triple terminology

In the setting of (S0,L2,S0

′) a quite similar results is due toGrochenig and coauthors:

Theorem

Assume that for some g ∈ S0 the Gabor frame operatorS : f 7→

∑λ∈Λ〈f , gλ〉gλ is invertible at the Hilbert space level, then

S defines automatically an automorphism of the BGT (S0,L2,S0

′).Equivalently, when g ∈ S0 generates a Gabor frame (gλ), then thedual frame (of the form (gλ)) is also generated by the elementg = S−1(g) ∈ S0.

The first version of this result has been based on matrix-valuedversions of Wiener’s inversion theorem, while the final result( Grochenig and Leinert, see [15]) makes use of the conceptof symmetry in Banach algebras and Hulanicki’s Lemma.



Classical Results in Banach Space Theory 2

Theorem (Theorem by S. Banach)

Assume that a linear mapping between two Banach spaces iscontinuous, and invertible as a mapping between sets, then it isautomatically an isomorphism of Banach spaces, i.e. the inversemapping is automatically linear and continuous.

So we have invertibility only in a more comprehensive category,and want to conclude invertibility in the given smaller (or richer)category of objects.



Other relations to the finite-dimensional case

The paper [14]: Gabor frames without inequalities Int. Math. Res.Not. IMRN, No.23, (2007) contains another collection ofstatements, showing the strong analogy between afinite-dimensional setting and the setting of Banach Gelfand triples:The main result (Theorem 3.1) of that paper shows, that theGabor frame condition (which at first sight looks just like atwo-sided norm condition) is in fact equivalent to injectivity of theanalysis mapping (however at the “outer level”, i.e. from S0

′(Rd )into `∞(Zd )), while it is also equivalent to surjectivity of thesynthesis mapping, but this time from `1(Zd ) onto S0(Rd ).



Kernel Theorem for general operators in L(S0,S0′)

Theorem

If K is a bounded operator from S0(Rd ) to S0′(Rd ), then there

exists a unique kernel k ∈ S0′(R2d ) such that 〈Kf , g〉 = 〈k , g ⊗ f 〉

for f , g ∈ S0(Rd ), where g ⊗ f (x , y) = g(x)f (y).

Formally sometimes one writes by “abuse of language”

Kf (x) =

∫Rd

k(x , y)f (y)dy

with the understanding that one can define the action of thefunctional Kf ∈ S0

′(Rd ) as

Kf (g) =

∫Rd

∫Rd

k(x , y)f (y)dy g(x)dx =

∫Rd

∫Rd

k(x , y)g(x)f (y)dxdy .



Kernel Theorem II: Hilbert Schmidt Operators

This result is the “outer shell” of the Gelfand triple isomorphism.The “middle = Hilbert” shell which corresponds to the well-knownresult that Hilbert Schmidt operators on L2(Rd ) are just thosecompact operators which arise as integral operators withL2(R2d )-kernels. The complete picture can be best expressed by aunitary Gelfand triple isomorphism. First the innermost shell:

Theorem

The classical kernel theorem for Hilbert Schmidt operators isunitary at the Hilbert spaces level, with 〈T , S〉HS = trace(T ∗ S ′)as scalar product on HS and the usual Hilbert space structure onL2(R2d ) on the kernels. An operator T has a kernel inK ∈ S0(R2d ) if and only if the T maps S0

′(Rd ) into S0(Rd ),boundedly, but continuously also from w∗−topology into the normtopology of S0(Rd ).



Kernel Theorem III

Remark: Note that for such regularizing kernels in K ∈ S0(R2d )the usual identification. Recall that the entry of a matrix an,k isthe coordinate number n of the image of the n−th unit vectorunder that action of the matrix A = (an,k ):

k(x , y) = T (δy )(x) = δx (T (δy )).

Note that δy ∈ S0′(Rd ) implies that K (δy ) ∈ S0(Rd ) by the

regularizing properties of K , hence pointwise evaluation is OK.With this understanding our claim is that the kernel theoremprovides a (unitary) isomorphism between the Gelfand triple (ofkernels) (S0,L

2,S0′)(R2d ) into the Gelfand triple of operator spaces

(L(S0′,S0),HS,L(S0,S0

′)).



AN IMPORTANT TECHNICAL warning!!

How should we realize these various BGT-mappings?Recall: How can we check numerically that e2πi = 1??Note: we can only do our computations (e.g. multiplication,division etc.) properly in the rational domain Q, we get to R byapproximation, and then to the complex numbers applying “thecorrect rules” (for pairs of real numbers).In the BGT context it means: All the (partial) Fourier transforms,integrals etc. only have to be meaningful at the S0-level (usingsimply Riemanian integrals!), typically isometric in the L2-sense,and extend by duality considerations to S0

′ when necessary, usingw∗-continuity!The Fourier transform is a good example (think of Fourierinversion and summability methods), similar arguments applyto the transition from the integral kernel of a linearmapping to its Kohn-Nirenberg symbol., for example.



Frames and Riesz Bases: the Diagram

P = C ◦R is a projection in Y onto the range Y0 of C, thus wehave the following commutative diagram.

Y

X Y0-C

� R ?

P

��

��

R



The frame diagram for Hilbert spaces:

`2(I )

H C(H)-C

� R ?

P

��

��

R



The frame diagram for Hilbert spaces (S0,L2,S0

′):

(`1, `2, `∞)

(S0,L2,S0

′) C((S0,L2,S0

′))-C

� R ?

P

��

��

R



Verbal Description of the Situation

Assume that g ∈ S0(Rd ) is given and some lattice Λ. Then (g ,Λ)generates a Gabor frame for H = L2(Rd ) if and only if thecoefficient mapping C from (S0,L

2,S0′)(Rd ) into (`1, `2, `∞)(Λ) as

a left inverse R (i.e. R ◦ C = IdH ), which is also aGTR-homomorphism back from (`1, `2, `∞) to (S0,L

2,S0′).

In practice it means, that the dual Gabor atom g is also in S0(Rd ),and also the canonical tight atom S−1/2, and therefore the wholeprocedure of taking coefficients, perhaps multiplying them withsome sequence (to obtain a Gabor multiplier) and resynthesis iswell defined and a BGT-morphism for any such pair.



Gabor frame matrix representations

Much in the same way as basis in Cn are used in order to describelinear mappings as matrices we can also use Gabor frameexpansions in order to describe (and analyze resp. betterunderstand) certain linear operators T (slowly variant channels,operators in Sjoestrand’s class, connected with another family ofmodulation spaces) by their frame matrix expansion. Working (forconvenience) with a Gabor frame with atom g ∈ S0(Rd ) (e.g.Gaussian atom, with Λ = aZ× bZ), and form for λ, µ ∈ Λ theinfinite matrix

aλ,µ := [T (π(λ)g)](π(µ)g).

This makes sense even if T maps only S0(Rd ) into S0′(Rd )!



Gabor frame matrix representations II

For any good Gabor family (tight or not) the mappingT 7→ A = (aλ,µ) is it self defining a frame representation, hence aretract diagram, from the operator BGT (B,H,B′) into the(`1, `2, `∞) over Z2d !In other words, we can recognize whether an operator isregularizing, i.e. maps S0

′(Rd ) into S0(Rd ) (with w∗-continuity) ifand only if the matrix has coefficients in `1(Z2d ).Note however, that invertibility of T is NOT equivalent toinvertibility of A! (one has to take the pseudo-inverse).



The Spreading Representation

The kernel theorem corresponds of course to the fact that everylinear mapping T from Cn to Cn can be represented by a uniquelydetermined matrix A, whose columns ak are the images T (~ek ).When we identify CN with `2(ZN) (as it is suitable wheninterpreting the FFT as a unitary mapping on CN) there is anotherway to represent every linear mapping: we have exactly N cyclicshift operators and (via the FFT) the same number of frequencyshifts, so we have exactly N2 TF-shifts on `2(ZN). They even forman orthonormal system with respect to the Frobenius norm,coming from the scalar product

〈A,B〉Frob :=∑k,j

ak,j bk,j = trace(A ∗ B ′)

This relationship is called the spreading representation of thelinear mapping T resp. of the matrix A. It can be thought asa kind of operator version of the Fourier transform.



The unitary spreading BGT-isomorphism

Theorem

There is a natural (unitary) Banach Gelfand triple isomorphism,called the spreading mapping, which assigns to operators T from(B,H,B′) the function or distribution η(T ) ∈ (S0,L

2,S0′)(R2d ).

It is uniquely determined by the fact that T = π(λ) = MωTt

corresponds to δt,ω.

Via the symplectic Fourier transform, which is of course anotherunitary BGT-automorphism of (S0,L

2,S0′)(R2d ) we arrive at the

Kohn-Nirenberg calculus for pseudo-differential operators. In otherwords, the mapping T 7→ σT = F sympη(T ) is another unitaryBGT isomorphism (onto (S0,L

2,S0′)(R2d ), again).



Consequences of the Spreading Representation

The analogy between the ordinary Fourier transform for functions(and distributions) with the spreading representation of operators(from nice to most general within our context) has interestingconsequences.We know that Λ-periodic distributions are exactly the ones havinga Fourier transform supported on the orthogonal lattice Λ⊥, andperiodizing an L1-function corresponds to sampling its FT.For operators this means: an operator T commutes with alloperators π(Λ), for some Λ C Rd × Rd , if and only ifsupp(η(T )) ⊂ Λ◦, the adjoint lattice. The Gabor frame operator isthe Λ-periodization of Pg : f 7→ 〈f , g〉g , hence η(S) is obtained bymultiplying η(Pg ) = Vg (g) pointwise by tt Λ◦ =

∑λ◦∈Λ◦ δλ◦ .



Consequences of the Spreading Representation 2

This observation is essentially explaining the Janssenrepresentation of the Gabor frame operator (see [7]).Another analogy is the understanding that there is a class ofso-called underspread operators, which are well suited to modelslowly varying communication channels (e.g. between the basisstation and your mobile phone, while you are sitting in the - fastmoving - train).These operators have a known and very limited support of theirspreading distributions (maximal time- and Doppler shift on thebasis of physical considerations), which can be used to “sample”the operator (pilot tones, channel identification) andsubsequently decode (invert) it (approximately).



Summability of sequences and quality of operators

One can however also fix the Gabor system, with both analysis andsynthesis window in S0(Rd ) (typically one will take g and grespectively, or even more symmetrically a tight Gabor window).Then one can take the multiplier sequence in different sequencespaces, e.g. in (`1, `2, `∞)(Λ).

Lemma

Then the mapping from multiplier sequences to Gabor multipliersis a Banach Gelfand triple homomorphism into Banach Gelfandtriple of operator ideals, consisting of the Schatten classe S1 =trace class operators, H = HS, the Hilbert Schmidt operators, andthe class of all bounded operators (with the norm and strongoperator topology).



Automatic continuity (> Balian-Low)

In contrast to the pure Hilbert space case (the box-function is anideal orthonormal system on the real line, but does NOT allow forany deformation, without loosing the property of being even aRiesz basis):

Theorem (Fei/Kaiblinger, TAMS)

Assume that a pair (g ,Λ), with g ∈ S0(Rd ) defines a Gabor frameor a Gabor Riesz basis respectively [note that by Wexler/Raz andRon/Shen these to situations are equivalent modulo taking adjointsubgroups!], then the same is true for slightly perturbed atoms orlattices, and the corresponding dual atoms (biorthogonalgenerators) depend continuously in the

(S0(Rd ), ‖ · ‖S0

)-sense on

both parameters.



THANK YOU!

Thank you for your attention!

Most of the referred papers of NuHAG can be downloaded fromhttp://www.univie.ac.at/nuhag-php/bibtex/

Furthermore there are various talks given in the last few years onrelated topics (e.g. Gelfand triples), that can be found bysearching by title or by name inhttp://www.univie.ac.at/nuhag-php/nuhag talks/



Selection of bibliographic items, see www.nuhag.eu

J. J. Benedetto and W. Czaja.

Integration and Modern Analysis.Birkhauser, 2009.

H. Bolcskei, F. Hlawatsch, and H. G. Feichtinger.

Equivalence of DFT Filter banks and Gabor expansions.In SPIE 95, Wavelet Applications in Signal and Image Processing III, volume 2569, part I, pages 128–139,San Diego, July 1995.

T. Chonavel.

Statistical Signal Processing. Modelling and Estimation. Transl. from the French by Janet Ormrod. Incl. 1CD-ROM.Springer, London, 2002.

I. Daubechies, A. Grossmann, and Y. Meyer.

Painless nonorthogonal expansions.J. Math. Phys., 27(5):1271–1283, May 1986.

R. J. Duffin and A. C. Schaeffer.

A class of nonharmonic Fourier series.Trans. Amer. Math. Soc., 72:341–366, 1952.

H. G. Feichtinger.

On a new Segal algebra.Monatsh. Math., 92:269–289, 1981.

H. G. Feichtinger and W. Kozek.

Quantization of TF lattice-invariant operators on elementary LCA groups.In H. G. Feichtinger and T. Strohmer, editors, Gabor analysis and algorithms, Appl. Numer. Harmon. Anal.,pages 233–266. Birkhauser Boston, Boston, MA, 1998.

H. G. Feichtinger and D. Onchis.

Constructive realization of dual systems for generators of multi-window spline-type spaces.J. Comput. Appl. Math., 234(12):3467–3479, 2010.

M. Frazier and B. Jawerth.

Decomposition of Besov spaces.Indiana Univ. Math. J., 34:777–799, 1985.

M. Frazier and B. Jawerth.

A discrete transform and decompositions of distribution spaces.J. Funct. Anal., 93(1):34–170, 1990.

L. Grafakos.

Classical and Modern Fourier Analysis.Prentice Hall, 2004.

J. Griffin and K. McKennon.

Multipliers and the group lp -algebras.Pacific J. Math., 49(2):365–370, 1973.

K. Grochenig.

Describing functions: atomic decompositions versus frames.Monatsh. Math., 112(3):1–41, 1991.

K. Grochenig.

Gabor frames without inequalities.Int. Math. Res. Not. IMRN, (23):Art. ID rnm111, 21, 2007.

K. Grochenig and M. Leinert.

Wiener’s lemma for twisted convolution and Gabor frames.J. Amer. Math. Soc., 17:1–18, 2004.

P. Grohs.

Continuous shearlet tight frames.J. Fourier Anal. Appl., 17(3):506–518, 2011.

C. Heil.

A Basis Theory Primer. Expanded ed.Applied and Numerical Harmonic Analysis. Basel: Birkhauser, 2011.

E. Hewitt and K. A. Ross.

Abstract Harmonic Analysis. Vol. II: Structure and Analysis for Compact Groups. Analysis on LocallyCompact Abelian Groups.Springer, Berlin, Heidelberg, New York, 1970.

J. Jost.

Postmodern Analysis. 2nd ed.Springer, Berlin, 1999.

Y. Katznelson.

An Introduction to Harmonic Analysis. 2nd corr. ed.Dover Publications Inc, New York, 1976.

R. Lasser.

Introduction to Fourier Series.Marcel Dekker, New York, NY, 1996.

R. Marks.

Handbook of Fourier Analysis and its Applications.Oxford University Press, 2009.

H. Reiter and J. D. Stegeman.

Classical Harmonic Analysis and Locally Compact Groups. 2nd ed.Clarendon Press, Oxford, 2000.

R. Shakarchi and E. M. Stein.

Fourier Analysis: An Introduction.Princeton Lectures in Analysis. Princeton University Press, 2003.

H. J. Weaver.

Applications of Discrete and Continuous Fourier Analysis.Wiley-Interscience, 1983.

M. Wong.

Discrete Fourier analysis.Pseudo-Differential Operators. Theory and Applications 5. Basel: Birkhauser. viii, 176 p., 2011.

A. Zygmund.

Trigonometric series. 2nd ed. Vols. I, II.Cambridge University Press, New York, 1959.



Invertibility, Surjectivity and Injectivity

In another, very recent paper, Charly Groechenig has discoveredthat there is another analogy to the finite dimensional case: Thereone has: A square matrix is invertible if and only if it is surjectiveor injective (the other property then follows automatically).We have a similar situation here (systematically describe inCharly’s paper):K.Groechenig: Gabor frames without inequalities,

Int. Math. Res. Not. IMRN, No.23, (2007).



Matrix-representation and kernels

We know also from linear algebra, that any linear mapping can beexpressed by a matrix (once two bases are fixed). We have asimilar situation through the so-called kernel theorem. It usesB = L(S0

′,S0).

Theorem

There is a natural BGT-isomorphism between (B,H,B′) and(S0,L

2,S0′)(R2d ). This in turn is isomorphic via the spreading and

the Kohn-Nirenberg symbol to (S0,L2,S0

′)(Rd × Rd ). Moreover,the spreading mapping is uniquely determined as theBGT-isomorphism, which established a correspondence betweenTF-shift operators π(λ) and the corresponding point masses δλ.



The w ∗− topology: a natural alternative

It is not difficult to show, that the norms of (S0,L2,S0

′)(Rd )correspond to norm convergence in (L1,L2,L∞)(R2d ).Therefore it is interesting to check what the w∗-convergence lookslike:

Lemma

For any g ∈ S0(Rd ) a sequence σn is w∗-convergent to σ0 if andonly the spectrograms Vg (σn) converge uniformly over compactsets to the spectrogram Vg (σ0).

The FOURIER transform, viewed as a BGT-automorphism isuniquely determined by the fact that it maps pure frequenciesonto the corresponding point measures δω.



A Typical Musical STFT



The w ∗− topology: dense subfamilies

From the practical point of view this means, that one has to lookat the spectrograms of the sequence σn and verify whether theylook closer and closer the spectrogram of the limit distributionVg (σ0) over compact sets.The approximation of elements from S0

′(Rd ) takes place by abounded sequence.Since any Banach-Gelfand triple homomorphism preserves thisproperty (by definition) one can reduce many problems tow∗-dense subsets of

(S0(Rd ), ‖ · ‖S0

).

Let us look at some concrete examples: Test-functions, finitediscrete measures µ =

∑i ciδti , trigonometric polynomials

q(t) =∑

i ai e2πiωi t , or discrete AND periodic measures

(this class is invariant under the generalized Fourier transformand can be realized computationally using the FFT).



The w ∗− topology: approximation strategies

How to approximate general distributions by test functions:Regularization procedures via product convolution operators,hα(gβ ∗ σ)→ σ or TF-localization operators: multiply theSTFT with a 2D-summability kernel before resynthesis (e.g.partial sums for Hermite expansion);

how to approximate an L1-Fourier transform by test functions:and classical summability

how to approximate a test function by a finite disretesequence using quasi-interpolation (N. Kaiblinger):QΨf (x) =

∑i f (xi )ψi (x).


ideas for a postmodern harmonic analysis · numerical harmonic analysis group ideas for a...

Documents